超宽带滤波器的稀疏贝叶斯正则化逆向神经网络建模  

SPARSE BAYESIAN REGULARIZED INVERSE NEURAL NETWORK MODELING OF UWB FILTER

作　　者：南敬昌[1] 王梓琦 高明明[1] 王颖[1] Nan Jingchang;Wang Ziqi;Gao Mingming;Wang Ying(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,Liaoning,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105

出　　处：《计算机应用与软件》2018年第10期232-237,共6页Computer Applications and Software

基　　金：国家自然科学基金项目(61372058);辽宁省教育厅重点实验室项目(LJZS007);辽宁省教育厅科学研究一般项目(L2015209)

摘　　要：针对射频器件建模中使用直接逆向神经网络精度较低,BP逆向神经网络泛化能力较差的问题,提出一种性能函数为贝叶斯L1/2范数的逆向神经网络建模方法。贝叶斯方法调整网络权系数避免过拟合现象,使模型输出更加平滑;增加L1/2范数扩充输入向量,使网络结构稀疏化且泛化能力更强。应用于超宽带滤波器谐振器逆向建模中,根据陷波频率处插入损耗值,求解对应的长度和宽度。结果表明:该方法与BP逆向建模方法相比,求得的长度、宽度和频率相对误差分别减小81. 4%、99. 8%、48. 9%,网络运行时间减少16. 3%,不存在多解问题,建模效率更高。In the modeling of radio freequency devices, direct inverse neural network has low precision and BP inverse neural network has poor generalization ability. To solve these problems, we proposed an inverse neural network modeling method with Bayesian L1/2 noml. The Bayesian method adjusted the network weight coefficient to avoid the over-fitting phenomenon so that the model output was much smoother. The input vector was expanded by adding the L1/2 noml so that the network structure was sparse and the generalization ability was much stronger. The method was applied into the inverse modeling of the UWB filter resonator. The value of the loss was inserted according to the notch frequency and the corresponding length and width were solved. The results show that compared with the BP inverse modeling method, the relative errors of length, width, and frequency obtained by this method are respectively reduced by 81.4% , 99.8% , and 48.9%. And the network operation time is decreased by 16.3%. This method does not have muhiple solutions and has better modeling efficiency.

关 键 词：神经网络 逆向建模 贝叶斯 L1/2范数 超宽带滤波器 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]